A method based on the successful AutoClass (Cheeseman & Stutz, 1996) and the Snob
research programs (Wallace & Dowe, 1994); (Baxter & Wallace, 1996) has been chosen for
our research work on harmonic classification. The method utilizes mixture models
(McLachlan, 1992) as a representation of the formulated clusters. This research is principally
based on the formation of such mixture models (typically based on Gaussian distributions)
through a Minimum Message Length (MML) encoding scheme (Wallace & Boulton, 1968).
During the formation of such mixture models the various derivative tools (algorithms) allow for the automated selection of the number of clusters and for the calculation of means,
variances and relative abundance of the member clusters. In this work a novel technique has
been developed using the MML method to determine the optimum number of clusters (or
mixture model size) during the clustering process. Once the optimum model size is
determined, a supervised learning algorithm is employed to identify the essential features of
each member cluster, and to further utilize these in predicting which ideal clusters any new
observed data may best described by.
This chapter first describes the design and implementation of the harmonic monitoring
program and the data obtained. Results from the harmonic monitoring program using both
unsupervised and supervised learning techniques are then analyzed and discussed.